FSD may require a hardware upgrade...

[3Victoria]

You guys are so absolute in your ideas ... amazing.
1. AFAIK California reports of testing only happen yearly, so too soon to make any comment on Tesla testing on public roads in California. Even if it is zero, that doesn't mean other testing on private or other states is not happening.
2. Mobileye wanted to go slower, and to control the data. Tesla wanted control. Hence a breakup. But the plan was to overlap the new with the old, and Tesla was forced to reconstruct the Mobileye functionality. By reports on TMC, I would say AP1 and AP2 are very close. (And yes, automatic wipers is a not important.)
3. My read on the demo videos is that most of that testing was on the integration of the NN for driving with the mapping and route setting software. The disengagements were largely for route failures. Eg:
5YJXCCE43GF017809 10/16/2016 < 0.2 Pl anner Output Invalid Unknown Wet Road
5YJXCCE43GF017809 10/16/2016 0.3 Follower Output Invalid Suburb an Road Wet Road
5YJXCCE43G F017809 10/16/2016 1.4 Follower Output Invalid Unknown Wet Road
The last demo showed no disengagements:
Nov 0 20 0 0
20 miles driven and no disengagements.
4. The demo videos were just that demos. They likely used an early NN model, and that was hacked to the mapper and route planner. So obviously a ways to go. How much is anyones *guess* at this point, unless you are on the dev team.

 

[MP3Mike]

He said it won't follow a fixed route and would react to traffic and congestion. Maybe that doesn't mean anything at all. We'll see.

He also said that you could pick almost any destination in the lower 48 at the last minute and it would be able to route and handle it. But the only way I could see that is if they upgraded all of the Superchargers to have an automated stall, or they take a truck with them that sets up at the SC just before the FSD vehicle gets there to charge. (I had assumed that they would setup automated SC stalls along the planned route, but maybe not.) Either that or they are just going to have a concierge/human helper plug/unplug the car at each SC.

 

[AnxietyRanger]

He said it won't follow a fixed route and would react to traffic and congestion. Maybe that doesn't mean anything at all. We'll see.

Sure, I'm not saying they'll program in a fixed set of steering inputs and press play. But still, training the car for a bunch of alternatives on the way to the other coast under a controlled test is quite different from a system that could be deployed everywhere, so we shall see how little or much useful stuff we can eventually extrapolate from that. Who knows, maybe they'll surprise us.

I do believe FSD codebase is separate and has not been seen in public since the video, so maybe they will surprise us. Maybe.

To set the record straight, I (like most here) get tired of the Negative Nellies that pounce on every word or announcement looking for something to parse and criticize.

I would say most don't care, some some vocal ones do. TMC has gotten a lot more mainstream where criticism of a company is not taken as personally. IMO that's a good thing. I for one don't mind at all that some are very critical, Tesla certainly has done a lot to deserve it too and the criticism (e.g. lawsuits) seems to make them a better company than the less critical cheering that lead to some unfortunate hubris IMO in the past...

 

[AnxietyRanger]

1. AFAIK California reports of testing only happen yearly, so too soon to make any comment on Tesla testing on public roads in California. Even if it is zero, that doesn't mean other testing on private or other states is not happening.

Possible.

2. Mobileye wanted to go slower, and to control the data. Tesla wanted control. Hence a breakup. But the plan was to overlap the new with the old, and Tesla was forced to reconstruct the Mobileye functionality. By reports on TMC, I would say AP1 and AP2 are very close. (And yes, automatic wipers is a not important.)

Yes, maybe and no.

Yes in the sense that Tesla did decide to move towards Drive PX. Yes in the sense that control of the data and speed of development probably were issues and Tesla was forced to reconstruct. Maybe in the sense that the Brown incident may have been a final straw for MobilEye.

No in the sense that AP1 and AP2 are very close. No, they are nowhere near very close. Even if we forget about the differences in functionality (e.g. AP2 doing to identifications on the screen), the confidence level for the driver is massively different. As in a different ballpark.

20 miles driven and no disengagements.

On a trained route, sure.

4. The demo videos were just that demos. They likely used an early NN model, and that was hacked to the mapper and route planner. So obviously a ways to go. How much is anyones *guess* at this point, unless you are on the dev team.

Yes, I agree it is a guess on how far they are. EAP does not tell us anything about that since it obviously is separate.

What might tell us something are the engineering personnel changes in AP as well as similar projects (and their length time-wise) by the competition. These might suggest trouble in Tesla's end, but who knows, maybe they will surprise.

 

[dknisely]

He said it won't follow a fixed route and would react to traffic and congestion. Maybe that doesn't mean anything at all. We'll see.

The promises of FSD are far and away my biggest hangup with Elon Musk. I'm a huge fan (and also a Tesla investor). But he has completely lost credibility with me on autonomy.

To set the record straight, I (like most here) get tired of the Negative Nellies that pounce on every word or announcement looking for something to parse and criticize. I also understand nearly all of his timelines are far too optimistic. But I feel his AP2.0 comments last year will haunt him in the future. He went a step too far IMO.

"Would react to traffic and congestion" is essentially untestable. OK, fine. How do I know you didn't plan to go that way anyway? Do I think it is likely that it will leave the interstate highway system and branch through some small town? Hardly. It might show the ability to navigate a good bit of the interstate highway system, including on/off ramps, but those are all well-charted and mapped.

Actually, the claim that they will have enough automated superchargers along the way is the most intriguing part, and would actually be the most significant demo if they deploy enough snaky chargers and they all work correctly.

 

[stopcrazypp]

Do you seriously believe that? I mean you really can't say you have the actual facts behind the production of that video.

That video was done almost a year ago at this point. If they were that far along then and there there were no disconnects at all over 20 miles of self driving, then where the heck is FSD or any partial features even resembling FSD??? It's certainly not held up in the hands of regulators.

Tesla's lack of communication here is telling and when they do communicate it is generally misleading or incomplete.

I'm only reporting what can be found in the report. It requires no inside knowledge, only knowing how to read and matching up the dates. You can double check the disengagement report if you want (linked in my original post).

As for a new video, we all know the next milestone being worked on is the cross country trip.

 

[stopcrazypp]

To my knowledge (definitely could be wrong), Tesla hasn't been doing any autonomous testing since 2016 (none filed with CA authorities or spotted in the wild), and the complete lack of any demos or even hints of demos seems like pretty strong evidence that they were using toy code (just like virtually every other self-driving project out there) to drive the same kind of route that looks cool, but isn't all that hard as long as you have a human supervising just in case a pedestrian or biker (read: actor) misses a cue. The MV demos all look alike for good reasons. That demo code isn't even .1% of what is required for L4, let alone L5.

Don't get me wrong -- I am a huge Tesla and Elon fan and knew exactly what I was getting into in December 2016, and I paid for EAP and FSD just as an incentive to get to solid L3 and maybe someday a little L4 on the highway. Obviously the L5 goal was merely a way to end up at L4 (maybe), while also learning what it will really take to do L5 with AP3 or AP4 or AP5. But my most pessimistic assessment never predicted AP2 far underperforming AP1 a year later or that Elon wouldn't find a way to make auto-wipers work without a rain sensor (at least in a crappy way!). The MobileEye situation was actually a catastrophe, and I sure hope Tesla can recover, but my expectations for AP2 cars are lower than AP1 at this point. AP2 may never drive as smoothly and confidently. [One exception -- AP2 does detect previously-untracked stopped cars that you come upon much better than AP1, and I'm happy for that. Vision >> radar for that purpose, and AP1 will never handle that dangerous case.]

You guys are so absolute in your ideas ... amazing.
1. AFAIK California reports of testing only happen yearly, so too soon to make any comment on Tesla testing on public roads in California. Even if it is zero, that doesn't mean other testing on private or other states is not happening.

Possible.

The disengagement reports are released only annually, and are not due until January of the following year (covering up to November 30). So we won't hear about them until January 2018 at the earliest (DMV also takes some time to put them on website).
"The California Autonomous Vehicle Testing Regulations require every manufacturer authorized to test autonomous vehicles on public roads to submit an annual report summarizing the disengagements of the technology during testing. This "disengagement report" is due by January 1 of each year. The first report covers the period from the date of issuance of the Testing Permit to November 30 of the following year. Subsequent reports cover the period of December 1 to November 30."
Testing of Autonomous Vehicles

Again, Tesla's next goal post is the cross country drive. I doubt they will be releasing more demos until then.

 

[brianman]

(And yes, automatic wipers is a not important.)

Going to have to disagree with you here. I live in the Pacific Northwest so perhaps I'm biased.

 

[stopcrazypp]

Going to have to disagree with you here. I live in the Pacific Northwest so perhaps I'm biased.

He likely meant "is not as important" (typo). But the automatic wipers are a bit of a different issue since AP1 used dedicated sensors, which AP2+ doesn't have.

 

[Swift]

I'm only reporting what can be found in the report. It requires no inside knowledge, only knowing how to read and matching up the dates. You can double check the disengagement report if you want (linked in my original post).

As for a new video, we all know the next milestone being worked on is the cross country trip.

I think your interpretation of the data in the report relative to how the video was made is highly debatable. Ultimately, it will come out, likely during discovery if this isn’t resolved in a reasonable timeframe (I.e. during the expected lifecycle of the lease/purchase).

As for a new video, no one wants another scripted, staged demo. People want transparency and communication.

 

[supratachophobia]

To my knowledge (definitely could be wrong), Tesla hasn't been doing any autonomous testing since 2016 (none filed with CA authorities or spotted in the wild), and the complete lack of any demos or even hints of demos seems like pretty strong evidence that they were using toy code (just like virtually every other self-driving project out there) to drive the same kind of route that looks cool, but isn't all that hard as long as you have a human supervising just in case a pedestrian or biker (read: actor) misses a cue. The MV demos all look alike for good reasons. That demo code isn't even .1% of what is required for L4, let alone L5.

Don't get me wrong -- I am a huge Tesla and Elon fan and knew exactly what I was getting into in December 2016, and I paid for EAP and FSD just as an incentive to get to solid L3 and maybe someday a little L4 on the highway. Obviously the L5 goal was merely a way to end up at L4 (maybe), while also learning what it will really take to do L5 with AP3 or AP4 or AP5. But my most pessimistic assessment never predicted AP2 far underperforming AP1 a year later or that Elon wouldn't find a way to make auto-wipers work without a rain sensor (at least in a crappy way!). The MobileEye situation was actually a catastrophe, and I sure hope Tesla can recover, but my expectations for AP2 cars are lower than AP1 at this point. AP2 may never drive as smoothly and confidently. [One exception -- AP2 does detect previously-untracked stopped cars that you come upon much better than AP1, and I'm happy for that. Vision >> radar for that purpose, and AP1 will never handle that dangerous case.]

Check fillings for Ohio. I've heard of testing going on there.

 

[supratachophobia]

I think your interpretation of the data in the report relative to how the video was made is highly debatable. Ultimately, it will come out, likely during discovery if this isn’t resolved in a reasonable timeframe (I.e. during the expected lifecycle of the lease/purchase).

As for a new video, no one wants another scripted, staged demo. People want transparency and communication.

They've always wanted that. But Tesla seems to think mystery is somehow beneficial to them in the long run. Which many would argue, isn't. Early transparency with them garnered much of the foundation favor that exists today. But they should be careful not to erode that.

 

[Swift]

They've always wanted that. But Tesla seems to think mystery is somehow beneficial to them in the long run. Which many would argue, isn't. Early transparency with them garnered much of the foundation favor that exists today. But they should be careful not to erode that.

Agreed. I'm just asking for transparency and communication for those that bought FSD. Radio silence is bad.

 

[buttershrimp]

Apart from the additional processing power, the new driver facing camera could be a way to ensure folks are watching the road or facial recognition to prevent theft and trying to take home ugly passengers late at night... "alert! FSD has been disabled until 6 am, please return to the bar."

 

[Canuck]

As for a new video, no one wants another scripted, staged demo.

Has anything come out about that video since it was released?

It looked so promising but now it makes no sense that they had it working that good when all this time later AP2.0 has major problems from what I've read -- since I don't have it myself. But I can't imagine it was fraudulently produced. That would involve too many people and someone would talk -- it always happens -- and I can't accept that Tesla is that stupid, or would cross that line. So I don't buy that it's scripted and staged even though that is possible it's highly improbable to me. Then again, I'd have never predicted that VW would make software to fool regulators when testing for emissions. I guess you can never underestimate the brazenness and stupidity when it comes to corporate greed.

So was it multiple takes pieced together where the car actually failed a lot, but when it did not the tape was cut and spliced together? I guess digitally you don't splice tape anymore but you know what I mean. They could argue that's not fraud, that the car did it, but to me while that may not be fraud it is scamming and unfortunately our advertising laws allow a lot of scamming.

In court you usually have to prove continuity of video, or have a very good reason for any edits, or judges will usually exclude it as unreliable. Plus, experts can tell if videos are edited and I would think Tesla's video has been reviewed by experts? Does anyone knows if it has been analyzed for edits? Or does anyone have some other explanation as to how they pulled it off?

 

[stopcrazypp]

They've always wanted that. But Tesla seems to think mystery is somehow beneficial to them in the long run. Which many would argue, isn't. Early transparency with them garnered much of the foundation favor that exists today. But they should be careful not to erode that.

Transparency also worked very badly for Tesla. I still remember the days when Tesla updated every step of the development for Model S and when inevitably they have to change things along the way, this caused huge problems (the controversy over the vanity mirrors and rear reading lights for example). This time with Model 3 they hid (and are still hiding) everything until the last minute, so that anything that is still subject to change can still change without setting false expectations. People seem to tend to take everything shown as a promise and ignore any caveats.

 

[Tjhappel]

As any AP2 owner will tell you there's literally no way that happened with any software or hardware we have/had/promised to have in a single shot. I can't drive in the carpool lane on the 405 which gets millions of tesla miles annually without having a disengagement over 5 miles, it's all over the place with lines etc and that is considered basic stiff for ap1 which I still drive in regularly. There is clearly zero fleet "learning" going on which is a real bummer.

Not sure what they did to get that video but my x isn't capable of driving safely on freeways with good markings let alone with pedestrians and stop signs etc...

 

[Swift]

Transparency also worked very badly for Tesla. I still remember the days when Tesla updated every step of the development for Model S and when inevitably they have to change things along the way, this caused huge problems (the controversy over the vanity mirrors and rear reading lights for example). This time with Model 3 they hid (and are still hiding) everything until the last minute, so that anything that is still subject to change can still change without setting false expectations. People seem to tend to take everything shown as a promise and ignore any caveats.

They need to be transparent and communicate with those that paid for FSD as well as those who paid for EAP (with the caveat that it would be delivered by 12/16). Not communicating at all on those items is the #1 reason they are getting beat up here. If I took money from my customers and went radio silent on them, I'm sure I'd be getting certified letters, bbb complaints, etc.

 

[lunitiks]

I't probably wasn't shot in a Hollywood studio like the moon landing was, but they might have some secret, artificial outdoor town that they did it in. Or the whole thing is cgi. I think I saw a lucasarts sign somewhere in the video, cant remember.

Ok kidding, but this is seriously where NN/DL ppl need to chime in. What IS actually a demo, and why is it different than production sw? What does it take to create a demo sw wise? And WHY oh why cant I create a demo for my daily route to work? Or the way I always park in my driveway? I'd accept that in a heartbeart and pay good mulah for it

 

[chillaban]

I't probably wasn't shot in a Hollywood studio like the moon landing was, but they might have some secret, artificial outdoor town that they did it in. Or the whole thing is cgi. I think I saw a lucasarts sign somewhere in the video, cant remember.

Ok kidding, but this is seriously where NN/DL ppl need to chime in. What IS actually a demo, and why is it different than production sw? What does it take to create a demo sw wise? And WHY oh why cant I create a demo for my daily route to work? Or the way I always park in my driveway? I'd accept that in a heartbeart and pay good mulah for it

Back when I was in college (MIT) I worked a little on our autonomous vehicle for the DARPA challenge. A popular strategy was to pre-map the challenge courses with extremely dense maps that were human-curated with plenty of manual measurements. At that point, the driving was essentially a script. You could paint the road pink and throw teddy bears everywhere and it would happily plow through all of that and think everything’s going according to plan.

At that point, you only have to implement some basic smarts that only kick in in limited scenarios: For example, stop at a stop sign by GPS, and then use vision to look for a gap in cross traffic. Heck you can even do a fairly shoddy job of that and you still won’t get into a wreck for a demonstration video: How many people do you know with flawed lane change algorithms who don’t get into accidents every time they drive?

In my last job, I was in the defense industry working almost exclusively on science fair projects, as we called them. It was all demoware, and we impressed a lot of customers and executives. It was possible to demonstrate features that looked intelligent but were nowhere near ready to be productized. But when you control the setup of the problem and the allowable inputs, demoware can look pretty damn good!


So, from that perspective, I’d say Tesla is actually operating in worst-case-scenarios with AP2/EAP right now. According to @verygreen’s tear downs, there’s basically zero prior knowledge being leveraged. It’s driving based off exactly what it sees, much like a human on a completely unfamiliar road. I would believe that Autosteer would work a ton better once they get their mapping system working, and they could really make it look artificially good by simply manually mapping some popular routes.